paper / santoroai / May 13 / failedTracing the Representation Geometry of Language Models from Pretraining to Post-training